Program Information
Effective Strategies for Deep Learning with Scarce Data for Prostate Lesion Classification Using Multiparametric MRI
D Karimi*, D Ruan , UCLA School of Medicine, Los Angeles, CA
Presentations
SU-K-601-5 (Sunday, July 30, 2017) 4:00 PM - 6:00 PM Room: 601
Purpose: Deep convolutional neural networks (CNNs) represent the state of the art in object detection and classification in computer vision. However, their application for detection and diagnosis of abnormal tissue in medical images is challenged by the scarcity of labeled training data. The goal of this study was to investigate the effectiveness of different CNN training strategies for classification of prostate cancer lesions in multi-parametric MRI from a small amount of data.
Methods: We adopted a transfer learning approach by fine-tuning a small and a large CNN, both pre-trained on large datasets of natural images, for prostate cancer lesion classification. Images of 232 lesions were available for training. We studied the effects of L1 and L2 regularization of the network weights and dropout on the classification performance. We also considered combining CNN-derived features with hand-crafted statistical features and texture features learned using a dictionary-based approach. Classification results are evaluated in terms of the area under the receiver operating characteristic curve (AUC).
Results: The small CNN achieved a higher AUC than the larger CNN (0.87 versus 0.80). L1 and L2 regularization of the network weights had a small effect on the classification performance, while using high dropout rates significantly improved the classification performance. Augmenting CNN-derived features with statistical and texture features significantly improved the classification results. The best classification result with an AUC of 0.87 was obtained with a dropout rate of 0.85 and combining the CNN-learned features with statistical and texture features.
Conclusion: In the absence of large labeled datasets, training of CNNs for prostate cancer lesion classification is challenging. Our study shows that in addition to transfer learning and data augmentation, two strategies are effective in improving the classification performance: 1) using large dropout rates, 2) combining the CNN-learned features with extra features that are hand-crafted or learned separately.
Contact Email: